In the semantic communication-enabled Industrial Internet of Things system (SemCom-IIoT), ensuring efficient task accomplishment encounters a critical issue: how to timely and accurately transmit and process semantic information of moderate amount with minimum resource cost, namely improving the efficiency of semantic information applied to the task. To overcome the problem, we provide a joint optimization of multi-dimensional resource allocation and semantic retention ratio selection in this paper. Specifically, a novel performance metric named efficiency of semantic information (EoSI) is defined to quantify the aforementioned efficiency. Further, considering the dynamics of the industrial environment, we exploit a deep reinforcement learning algorithm to maximize EoSI of the system by adaptively adjusting the allocation of computing resources, bandwidth resources and the selection of semantic retention ratio. Extensive simulation results demonstrate that our proposed approach can significantly enhance EoSI by up to 173% and 139% respectively in the context of a large number of devices and the strict task accuracy requirement, compared with other benchmarks.